Unitree Go2 Locomotion

This is my practice of using MuJoCo Playground for training Unitree Go2 using PPO with real-world deployment. The training part is largely based on MuJoCo Playground with actual proprioceptions and a minimal set of rewards.

This repository is aimed for beginners such as undergraduate and graduate students. Please feel free to contact me if you have questions.

Author: Ella Yixuan Qiu
Email: yixuan.qiu [at] rug.nl

Go2 environment structure

The Go2 implementation lives in mujoco_playground/_src/locomotion/go2/:

• base.py: common Go2 MJX environment setup, asset loading from mujoco_menagerie/unitree_go2, PD gain configuration, and sensor helpers.
• joystick.py: velocity-command tracking task used by Go2JoystickFlatTerrain and Go2JoystickRoughTerrain.
• go2_constants.py: XML paths, root body name, foot site/geom names, and sensor names.
• randomize.py: domain randomization for friction, mass/inertia, PD gains, and torso center-of-mass offsets.
• getup.py and handstand.py: additional Go2 recovery and balance tasks.
• xmls/: MJCF scene/model files for flat and rough terrain variants, including feet-only and full-collision models.

The joystick policy uses a 12-dimensional action in MuJoCo actuator order:

FL_hip, FL_thigh, FL_calf,
FR_hip, FR_thigh, FR_calf,
RL_hip, RL_thigh, RL_calf,
RR_hip, RR_thigh, RR_calf

The actor observation is one 45-dimensional frame consisting of local angular velocity, projected gravity, command, joint-position error, joint velocity, and previous action. With history_len=10, exported policies expect:

input:  obs      [1, 450]
output: actions  [1, 12]

Installation

conda create -n env_mujocoplayground python=3.12 -y  
conda activate env_mujocoplayground # make sure environment is always activated in following commands
cd leg-mujoco
pip install -e . # install Go2 customized MuJoCo Playground
# install required dependencies
pip install rsl-rl-lib wandb

Train policy

train-rsl-ppo --env_name=Go2JoystickFlatTerrain

Play policy

It is necessary to specify play_only, load_run_name, and checkpoint_num (optional).

train-rsl-ppo \
  --env_name=Go2JoystickFlatTerrain \
  --play_only \
  --load_run_name=<run_name> \
  --checkpoint_num=<number>

If checkpoint_num is not specified, the latest checkpoint is automatically chosen.

Export policy to ONNX

This is for Sim-to-Sim validation and real-world deployment, we use ONNX format to deploy.

train-rsl-ppo \
  --env_name=Go2JoystickFlatTerrain \
  --export_onnx_only \
  --load_run_name=<run_name> \
  --checkpoint_num=<number>

Real-world Deployment

The ROS 2/Unitree deployment code lives in deploy and runs exported ONNX policies for Go2JoystickFlatTerrain. See deploy/README.md for setup, configuration, sim-to-sim, and real-robot notes.

Acknowledgements

This work is based on and adapted from:

• google-deepmind/mujoco_playground, the upstream MuJoCo Playground project.
• aatb-ch/mujoco_playground/tree/go2, whose Go2 branch informed the Go2 environment structure.

License

This is essentially an engineering practice of MuJoCo Playground on Unitree Go2. For more information on licensing, please refer to the original MuJoCo Playground.